An intelligent multi-agent distributed battlefield via Multi-Token Message Passing

Abstract

The main focus of this paper is on issues related to communications, load balancing, resource sharing and allocation in a distributed computing system, as applied to battlefield simulation. We present a distributed architecture of a middleware and algorithms for communications, scheduling and load balancing. In the proposed distributed architecture the nodes use the Multi-Token Message Passing Interface (MPI) over the message based Myrinet communication system, namely the GM protocol, on a Gigabit Myrinet switch to communicate with each other. As we show here this brings in additional implementation challenges regarding (a) the safe usage of threads and MPI, (b) expanding MPI to support additional data-types specific to the sought battlefield simulation applications, and (3) modifying MPI to achieve equal distribution of load among nodes. The main task in load balancing is to keep load equally distributed among the computing nodes, and yet keep the notion of geographically distributed agents transparent. An additional challenge is posed by the fact that when the application is running, load balancing has to be continuously adapted- i.e. no run-time stopping is allowed to redistribute the load among nodes.